- See Section 2.235.
- Response parameter is acceleration for 5% damping.
- Only use spectral accelerations within passband of filter (1.25f
_{l}and f_{h}) where f_{l}is the low cut-off frequency and f_{h}is the high roll-off frequency. - Note that after 0.8s the number of records available for regression analysis starts to decrease rapidly and that after 4s there are few records available. Only conduct regression analysis up to 2.5s because for longer periods there are too few records to obtain stable results. Note that larger amplitude ground motions are better represented in the set for long-periods (> 1s).
- Find that logarithmic transformation may not be justified for nine periods (0.26, 0.28 and 0.44–0.65s) by using pure error analysis but use logarithmic transformation since it is justified for neighbouring periods.
- By using pure error analysis, find that for periods > 0.95s the null hypothesis of a magnitude-independent standard deviation cannot be rejected so assume magnitude-independent σ. Note that could be because magnitude-dependent standard deviations are a short-period characteristic of ground motions or because the distribution of data w.r.t. magnitude changes at long periods due to filtering.
- Find that different coefficients are significant at different periods so try changing the functional form to exclude insignificant coefficients and then applying regression again. Find that predicted spectra show considerable variation between neighbouring periods therefore retained all coefficients for all periods even when not significant.
- Note that smoothing could improve the reliability of long-period ground-motion estimates because they were based on less data but that smoothing is not undertaken since the change of weighted to unweighted regression at 0.95s means a simple function cannot fit both short- and long-period coefficients.